Ensemble Prediction of Time-to-Event Outcomes with Competing Risks: A Case-Study of Surgical Complications in Crohn's Disease

Abstract

Summary We develop a novel algorithm to predict the occurrence of major abdominal surgery within 5 years following Crohn's disease diagnosis by using a panel of 29 baseline covariates from the Swedish population registers. We model pseudo‐observations based on the Aalen–Johansen estimator of the cause‐specific cumulative incidence with an ensemble of modern machine learning approaches. Pseudo‐observation preprocessing easily extends all existing or new machine learning procedures for continuous data to right‐censored event history data. We propose pseudo‐observation‐based estimators for the area under the time varying receiver operating characteristic curve, for optimizing the ensemble, and the predictiveness curve, for evaluating and summarizing predictive performance.

Stacked Inverse Probability of Censoring Weighted Bagging: A Case Study In the InfCareHIV Register

“…Unlike the recently proposed methods of Sachs et al. (2019) using pseudo‐observations, which although possibly more computationally efficient, requires that censoring is not only independent of the event time, but independent of all covariate information.…”

“…Due to the use of IPCW in both the bagging and the loss function, our proposed method can easily handle dependent censoring under coarsening at random. Unlike the recently proposed methods of Sachs et al (2019) using pseudo-observations, which although possibly more computationally efficient, requires that censoring is not only independent of the event time, but independent of all covariate information.…”

“…Sachs et al. (2019) used a pseudo‐observation based version of the time‐dependent AUC for this purpose.…”

Prediction meets causal inference: the role of treatment in clinical prediction models

“…Without covariates and without censoring for other reasons like loss to follow up or end of study, cumulative incidence is estimable by the number of patients with the event of interest divided by the total number of patients at baseline. With covariates and censoring for reasons like loss to follow up or end of study, the estimation can be done in various ways, see for instance [ 10 , 21 – 24 ]. Continued follow up after treatment initiation is not needed.…”

A Deep CNN Approach For Predicting Cumulative Incidence Based On Pseudo-Observations

“…Recent works in other areas of machine learning have suggested data pre-processing steps that can be used to adapt common classes of machine learning methods to time-to-event outcomes. Pseudo-observations [13] is one of such methods that has been suggested to adapt random forests [14][15] and more generally all methods for continuous outcomes and ensembles of them [16]. [17] proposed the use of a modified version of pseudo-observations of [13], which they call conditional, to replace the observed survival times to make risk predictions in deep neural network.…”